448 four-way questions: pick the funny one-character edit of a movie or band name; BIG-bench Hard keeps 250.
unassessed
| Category | reasoning |
|---|---|
| Subcategory | BIG-bench (and BBH) humorous one-edit movie and band names |
| Page status | unknown |
| Metric | multiple_choice_grade |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 448 |
| Dataset licence | Apache-2.0 |
| Publisher | Google (BIG-bench collaboration) |
ruin_names shows a movie or artist name and four near-edits. One edit is a humorous ruin from Twitter hashtags #RuinAFilmTitle and #RuinABandName; the others are non-funny Levenshtein-1 distractors. Authors Ryan Teehan, Vinay Prabhu, and Nick Roberts. The skill is spotting the joke, not generating names. BIG-bench Hard keeps a 250-item sample.
Multiple choice with preferred metric multiple_choice_grade. append_choices_to_input is true. 447 items have four options; one item has three. Dummy-model header: 448 multiple-choice queries. Canary GUID embedded.
No model card in ModelSpec reports this benchmark yet.